The model is the cheapest part of most AI products, and it is the only part anyone budgets for. The expensive parts are getting your data into a state a model can use, building the retrieval layer that finds the right piece of it, and running the evaluation that keeps the whole thing honest.

Data preparation is the first surprise. Documents that a human can read are not documents a model can search — they need cleaning, chunking, embedding and re-embedding every time the source changes. For most businesses this is weeks, not days.

Then there is the running cost, which unlike a normal application scales with usage rather than with users. A feature that costs a fraction of a cent per call is free at a hundred calls a day and a real line item at a hundred thousand. Model that curve before you ship, not after.

The rule of thumb we use: whatever the model work costs, plan for two to three times that around it. Teams that budget the other way round are the ones that stall at the pilot.